Accurate perception of scene geometry is a critical aspect of enabling Internet of Video Things (IoVT) devices to perform environmental perception and interaction. Monocular depth estimation has garnered widespread attention due to its configuration simplicity. However, existing methods often struggle to fully exploit the uniqueness and complementarity between single-frame and multi-frame cues, resulting in degraded accuracy and robustness in dynamic environments. In this paper, we propose a multi-frame monocular depth estimation method based on Dual-Branch Cost Volume Fusion (DB-CVF) in dynamic environments, which addresses the aforementioned issues through the unique representation and complementary fusion of cost volumes. DB-CVF represents the single-frame and multi-frame cues as multi-scale self-cost volume and attention-aware cross-cost volume, respectively, and achieves complementary fusion through a Multi-Cue Cost Volume Fusion (MC-CVF) module. A Single-Frame Cue Generation (SFCG) module and an Attention-Aware Multi-Frame Cue Generation (A2-MFCG) module are designed in a dual-stream structure to separately generate two types of cost volumes within the same latent space. The two types of cost volumes, each with its own uniqueness, realize information interaction and mutual benefit by the designed MC-CVF module based on the cross-attention mechanism. Extensive quantitative and qualitative results demonstrate that DB-CVF achieves outstanding performance on the KITTI, DDAD, and Cityscapes datasets. Compared to the state-of-the-art method, DB-CVF reduces the depth error in dynamic regions by 17.8%, showcasing the broad potential of cue fusion for monocular depth estimation.

Multimodal aerial image semantic segmentation enables fine-grained land cover classification by integrating data from different sensors, yet it remains challenged by information redundancy, inter-modal feature discrepancies, and class confusion in complex scenes. To address these issues, we propose a Cross-Modal Hierarchical Feature Fusion Network (CMHFNet) based on an encoder-decoder architecture. The encoder incorporates a Pixel-wise Attention-guided Fusion Module (PAFM) and a Multi-stage Progressive Fusion Transformer (MPFT) to suppress redundancy and model long-range intermodal dependencies and scale variations. The decoder introduces a residual information-guided feature compensation mechanism to recover spatial details and mitigate class ambiguity. Experiments on DDOS, Vaihingen, and Potsdam datasets demonstrate that CMHFNet surpasses state-of-the-art methods, validating its effectiveness and practical value.

Due to the absence of detailed information, such as texture, shape, and color, detecting infrared small targets remains a challenging problem. While existing model-driven and data-driven approaches have made some progress, they still struggle to effectively exploit global contextual information and frequency-specific details. In this paper, we introduce a Frequency-Aware Contextual Feature Pyramid Network (FACFPNet) to address these limitations in infrared small target detection. Specifically, we first estimate the correlation between high- and low-frequency feature representations within an encoder-decoder framework based on the ResNet-18 backbone. This is achieved through the Contextual Fine-Grained Block (CFGB), which effectively combines local fine-grained features with global semantic information for enhanced contextual feature modeling. Next, we propose a Frequency-Aware Attention Module (FAAM) to address the underutilization of prior frequency knowledge in infrared small targets, thereby improving the preservation of these features. This module enhances global contextual representation by more effectively extracting high- and low-frequency information. Finally, during the decoding stage, shallow fine-structure information is interactively fused with deep semantic features through the Asymmetric Enhancement Fusion module (AEFM), which strengthens the representation of small targets and improves information retention. Experimental results on three publicly available datasets, SIRST-Aug, MdvsFA and IRSTD-1K, demonstrate that our method achieves superior detection performance.

Self-supervised visual odometry (VO) has exhibited remarkable benefits over supervised methods, surpassing the reliance on the annotated ground-truth of training data. However, most existing self-supervised VO methods, namely scene appearance-based methods, have limitations in exploiting the complementary properties of cross-modal information between scene appearance and structure. To this end, we propose a novel self-supervised VO based on scene appearance-structure incremental fusion scheme. Specifically, a Global-Local Context awareness-based Depth estimation Network (GLC-DN) is designed to introduce the scene structural cues, thus laying the foundation for realizing the scene appearance-structure incremental fusion. Then, a Dual stream Pose estimation Network based on Scene Appearance-Structure Incremental Fusion (SASIF-DPN) is devised, which consists of a Dual Stream Network (DSN) and multiple Cross-Modal Complementary Fusion Modules (CM-CFMs). CM-CFM fully leverages the complementary properties between the RGB information and the predicted depth information, and the combination of multiple CM-CFMs facilitates the information interaction between the two modalities in an incremental fusion manner. Detailed evaluations of GLC-DN and SASIF-DPN provably confirm the effectiveness and design principles of each component we propose. Extensive comparison experiments have also been conducted, which clearly verify the superiority of our method compared to current counterparts.

Infrared Small Target Detection is a challenging task to separate small targets from infrared clutter background. Recently, deep learning paradigms have achieved promising results. However, these data-driven methods need plenty of manual annotations. Due to the small size of infrared targets, manual annotation consumes more resources and restricts the development of this field. This letter proposed a labor-efficient annotation framework with level set, which obtains a high-quality pseudo mask with only one cursory click. A variational level set formulation with an expectation difference energy functional is designed, in which the zero level contour is intrinsically maintained during the level set evolution. It solves the issue that zero level contour disappearing due to small target size and excessive regularization. Experiments on the NUAA-SIRST and IRSTD-1k datasets demonstrate that our approach achieves superior performance. Code is available at https://github.com/Li-Haoqing/COM.

In this paper, a deep learning (DL) framework was proposed to predict the taxi-passenger demand while the spatial, the temporal, and external dependencies were considered simultaneously. The proposed DL framework combined a modified density-based spatial clustering algorithm with noise (DBSCAN) and a conditional generative adversarial network (CGAN) model. More specifically, the modified DBSCAN model was applied to produce a number of sub-networks considering the spatial correlation of taxi pick-up events in the road network. And the CGAN model, fed with the historical taxi passenger demand and other conditional information, was capable to predict the taxi-passenger demands. The proposed CGAN model was made up with two long short-term memory (LSTM) neural networks, which are termed as the generative network G and the discriminative network D, respectively. Adversarial training process was conducted to the two LSTMs. In the numerical experiment, different model layouts were compared. It was found that different network layouts provided reasonable accuracy. With limited training data, more LSTM layers in the generator network resulted in not only higher accuracy, but also more difficulties in training. Comparisons were also conducted between the proposed prediction model and four typical approaches, including the moving average method, the autoregressive integrated moving method, the neural network model, and the LSTM neural network model. The comparison results showed that the proposed model outperformed all the other methods. And the repeated experiment indicated that the proposed CGAN model provided significant better predictions than the LSTM model did. Future research was recommended to include more datasets for testing the model and more information for improving predictive performance.

Salient object detection, as a necessary step of many computer vision applications, has attracted extensive attention in recent years. A novel salient object detection method is proposed based on multi‐superpixel‐scale contrast. Saliency value of each superpixel is measured with a global score, which is computed using the region's colour contrast and the spatial distances to all other regions in the image. High‐level information is also incorporated to improve the performance, and the saliency maps are fused across multiple levels to yield a reliable final result using the modified multi‐layer cellular automata. The proposed algorithm is evaluated and compared with five state‐of‐the‐art approaches on three publicly standard datasets. Both quantitative and qualitative experimental results demonstrate the effectiveness and efficiency of the proposed method.